evgltop 0.1.0

Interactive GPU monitoring TUI for SLURM clusters

See your GPUs. Know who's on them. Grab one in seconds.

---

The Problem

You SSH into the lab server. You need a GPU. You run nvidia-smi. All 4 GPUs are taken. But by who? Are they actually using them? When will one free up? You run squeue. You cross-reference job IDs. You check tmux sessions. You piece it together manually.

evgltop does all of this in one screen, updated every second.

Install

pip install evgltop

Then just run:

evgltop

That's it. No config, no setup, no root access.

• See all 4 GPUs at a glance with live utilization bars
• See exactly who is running what process on each GPU
• Press n to launch a new SLURM session with auto-optimized QOS
• Click Cancel to kill a job (and optionally its tmux session)
• Watch the pending queue with estimated wait times

---

Screenshots

GPU Dashboard + System Resources

Every GPU card shows: utilization, VRAM, temperature, power, the user, their SLURM job, runtime, and remaining time. Below: CPU, RAM, disk usage, and your tmux sessions.

One-Click Session Launch

Press n. Pick GPUs, partition, memory. evgltop auto-selects the best QOS so your job starts as fast as possible.

Pending Queue with Wait Time Estimates

All GPUs taken? evgltop shows your pending jobs with estimated start times. Even when SLURM says N/A, evgltop calculates it from running jobs' remaining time — and chains estimates across multiple pending jobs.

Built-In Help

Press h for a quick reference of everything.

---

Feature Highlights

GPU Waste Detection

If someone allocated a GPU but isn't using it (0% utilization for >10 min), the card turns red with IDLE status. Hard to ignore.

Smart QOS

SLURM QOS tiers have different GPU limits and priorities. evgltop checks what you're already using and picks the tier that maximizes your throughput:

Request 1 GPU:
  normal slot free?  -> normal  (priority 40000, fastest)
  normal full?       -> gpu02   (priority 20000)

Request 2 GPUs:      -> gpu02
Request 3-4 GPUs:    -> gpu04   (priority 10000)

No more QOSMaxGRESPerUser surprises.

Pending Queue with Time Estimates

SLURM says N/A for your pending job's start time? evgltop calculates it by looking at when running jobs will end, then chains the estimates if you have multiple pending jobs.

Job 64  gres:gpu:4  ~6d 23h (03/28 00:12)  QOS limit
Job 66  gres:gpu:4  ~7d 23h (03/29 00:12)  QOS limit   <- accounts for Job 64's runtime

Tmux Session Tracking

Sessions auto-named gpu{N}-{JOBID}. The Sessions panel shows which are running, waiting, or ended:

● gpu1-53    RUN  GPU 0  03/21 00:12   tmux attach -t gpu1-53
◌ gpu4-67    WAIT        03/21 00:36   tmux attach -t gpu4-67
○ gpu2-70    END         03/21 01:02   tmux attach -t gpu2-70

---

Keybindings

Key	Action
n	New SLURM session
h	Help
r	Force refresh
q	Quit
Esc	Close dialog
Click Cancel	Cancel a running/pending job

---

Requirements

• Python 3.10+
• NVIDIA GPUs with nvidia-smi
• SLURM (squeue, scontrol, scancel)
• tmux

---

Emory Vision & Graphics Lab